Perplexity's JOLT Makes Agent Training Dramatically Cheaper

Perplexity's JOLT Makes Agent Training Dramatically Cheaper

Perplexity, working with NVIDIA, released JOLT on October 21 — a post-training method where a single model plays both teacher and student during reinforcement learning [1]. The headline number: JOLT matches GRPO's accuracy on MATH benchmarks using 6.5x fewer tokens, and the upgraded JOLT+ actually beats GRPO outright on coding, tool-use, and terminal tasks, including a +9.9 point jump on AppWorld [2][3].

The technical trick is solving credit assignment — giving per-token feedback instead of a single outcome-level reward — which is the thing that makes RL training agents so expensive and slow in the first place. Aravind Srinivas framed this as a token-efficiency win, and the agent-builder crowd on X zeroed in on the same thing: cheaper, faster RL loops mean smaller labs can compete on agent quality without GPU budgets the size of a hyperscaler's.

This matters more than it sounds. If training efficient agents gets 6x cheaper, the moat shifts further away from "who has the most compute" and toward "who has the best data and the clearest judgment about what good behavior looks like." That's a direct hit on the compute-is-everything narrative.

Tesla's Q3 Shows the AI Roadmap Slipping — But the Market Doesn't Care Yet

Tesla posted Q3 2026 revenue of $28.05B, flat year-over-year, with non-GAAP EPS of $0.48 beating estimates and — notably — positive free cash flow despite heavy AI spend [1][2]. The AI story underneath the earnings: unsupervised FSD for customer-owned cars now targets Q2 2027 (previously sooner), the Robotaxi fleet hit roughly 2 million cumulative unsupervised miles (8x quarter-over-quarter growth), and Optimus high-volume production — aiming for ~20,000 units/week — is now a "~1 year" target with external deliveries sliding to late 2027 [3].

Translation: the timelines moved right again, as they always do, but the underlying metrics (mile growth, cash flow) are strong enough that analysts raised price targets anyway. Elon Musk and the Tesla team leaned hard into the 2027 narrative — Terafab chip commissioning, AI5 hardware — betting that investors will keep extending patience as long as the trendlines point up.

The real signal for builders: Tesla's Robotaxi mileage scaling 8x in a single quarter is the kind of compounding curve that's easy to dismiss until it isn't. Autonomy is still a hardware-plus-judgment problem, not a pure software one, and Tesla's slipping dates are a reminder that "the model works in the lab" and "the model works unsupervised at scale" remain very different bars.

OpenAI Doubles Down on Daybreak While Anthropic Takes Heat for Caution

OpenAI's expanded Daybreak program — now split into Blue and Red tiers — gives vetted security researchers access to frontier models like GPT-5.6-Cyber for real offensive and defensive work: exploit chains, vulnerability research, the works [1][2]. The compliance numbers are the story here — 95% task compliance from approved researchers versus 1.5-2% on standard consumer-facing models, meaning OpenAI built an entirely separate trust tier rather than loosening guardrails for everyone [3].

The framing in the research — and in the X discourse — explicitly contrasts this with Anthropic's more conservative safety classifiers, which researchers say block comparable vulnerability work outright. It's a genuinely interesting fork in safety philosophy: gatekeep access tightly but give real capability to the vetted few, versus apply blanket restrictions that are safer by default but frustrate legitimate security researchers.

Security researchers on X were vocal about preferring OpenAI's approach — not because they want fewer guardrails everywhere, but because "no" isn't a defensive cybersecurity strategy. As defenders race attackers who already have frontier models, this is a preview of a fight every AI lab will have to have: trust tiers aren't a nice-to-have, they're becoming infrastructure.

EU Parliament Fractures Over AI Act Liability Rules

The European Parliament spent October 22 debating AI Act implementation, and the coalition cracks are showing: center-right MEPs are pushing to simplify rules for competitiveness, while center-left and Green factions want to revive the AI Liability Directive to make harm accountability enforceable [1][2]. Risk-tier classifications and delays to high-risk category enforcement were central to the fight, alongside pressure to align with G7, UN, and US/China frameworks on cybersecurity and extraterritorial reach.

For Nordic AI builders, this is the fork in the road that actually matters more than any US product launch this week. If liability rules get watered down in the name of competitiveness, European companies get a lighter compliance load but less legal clarity about who's on the hook when an agent screws up a financial transaction or a medical recommendation. If the Liability Directive comes back with teeth, compliance gets harder but the rules of engagement get clearer — which, frankly, most serious builders would prefer over ambiguity.